floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

arm: Use assembly implementations based on the ARM abi 

In case of softfp abi, assembly implementations of only those APIs are
used which doesnt have a floating point argument or return value.

In case of hard abi, all assembly implementations are used.
This commit is contained in:
Ashwin Sekhar T K 2017-06-30 13:06:38 +05:30
parent da7f0ff425
commit 89924b3d5b
2 changed files with 60 additions and 177 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

104
kernel/arm/KERNEL.ARMV6
View File

@ -1,7 +1,5 @@
include $(KERNELDIR)/KERNEL.ARMV5
###############################################################################
SAMAXKERNEL = iamax_vfp.S SAMAXKERNEL = iamax_vfp.S
DAMAXKERNEL = iamax_vfp.S DAMAXKERNEL = iamax_vfp.S
CAMAXKERNEL = iamax_vfp.S CAMAXKERNEL = iamax_vfp.S
@ -34,6 +32,45 @@ IDMAXKERNEL = iamax_vfp.S
ISMINKERNEL = iamax_vfp.S ISMINKERNEL = iamax_vfp.S
IDMINKERNEL = iamax_vfp.S IDMINKERNEL = iamax_vfp.S
SGEMMKERNEL = ../generic/gemmkernel_4x2.c
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
SGEMMINCOPY = sgemm_ncopy_4_vfp.S
SGEMMITCOPY = sgemm_tcopy_4_vfp.S
SGEMMINCOPYOBJ = sgemm_incopy.o
SGEMMITCOPYOBJ = sgemm_itcopy.o
endif
SGEMMONCOPY = sgemm_ncopy_2_vfp.S
SGEMMOTCOPY = ../generic/gemm_tcopy_2.c
SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o
DGEMMKERNEL = ../generic/gemmkernel_4x2.c
ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
DGEMMINCOPY = dgemm_ncopy_4_vfp.S
DGEMMITCOPY = dgemm_tcopy_4_vfp.S
DGEMMINCOPYOBJ = dgemm_incopy.o
DGEMMITCOPYOBJ = dgemm_itcopy.o
endif
DGEMMONCOPY = dgemm_ncopy_2_vfp.S
DGEMMOTCOPY = ../generic/gemm_tcopy_2.c
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o
STRMMKERNEL = ../generic/trmmkernel_4x2.c
DTRMMKERNEL = ../generic/trmmkernel_4x2.c
CGEMMONCOPY = cgemm_ncopy_2_vfp.S
CGEMMOTCOPY = cgemm_tcopy_2_vfp.S
CGEMMONCOPYOBJ = cgemm_oncopy.o
CGEMMOTCOPYOBJ = cgemm_otcopy.o
ZGEMMONCOPY = zgemm_ncopy_2_vfp.S
ZGEMMOTCOPY = zgemm_tcopy_2_vfp.S
ZGEMMONCOPYOBJ = zgemm_oncopy.o
ZGEMMOTCOPYOBJ = zgemm_otcopy.o
ifeq ($(ARM_ABI),hard)
SASUMKERNEL = asum_vfp.S SASUMKERNEL = asum_vfp.S
DASUMKERNEL = asum_vfp.S DASUMKERNEL = asum_vfp.S
CASUMKERNEL = asum_vfp.S CASUMKERNEL = asum_vfp.S
@ -44,11 +81,6 @@ DAXPYKERNEL = axpy_vfp.S
CAXPYKERNEL = axpy_vfp.S CAXPYKERNEL = axpy_vfp.S
ZAXPYKERNEL = axpy_vfp.S ZAXPYKERNEL = axpy_vfp.S
SCOPYKERNEL = copy.c
DCOPYKERNEL = copy.c
CCOPYKERNEL = zcopy.c
ZCOPYKERNEL = zcopy.c
SDOTKERNEL = sdot_vfp.S SDOTKERNEL = sdot_vfp.S
DDOTKERNEL = ddot_vfp.S DDOTKERNEL = ddot_vfp.S
CDOTKERNEL = cdot_vfp.S CDOTKERNEL = cdot_vfp.S
@ -64,11 +96,6 @@ DROTKERNEL = rot_vfp.S
CROTKERNEL = rot_vfp.S CROTKERNEL = rot_vfp.S
ZROTKERNEL = rot_vfp.S ZROTKERNEL = rot_vfp.S
SSCALKERNEL = scal.c
DSCALKERNEL = scal.c
CSCALKERNEL = zscal.c
ZSCALKERNEL = zscal.c
SSWAPKERNEL = swap_vfp.S SSWAPKERNEL = swap_vfp.S
DSWAPKERNEL = swap_vfp.S DSWAPKERNEL = swap_vfp.S
CSWAPKERNEL = swap_vfp.S CSWAPKERNEL = swap_vfp.S
@ -90,57 +117,8 @@ CTRMMKERNEL = ctrmm_kernel_2x2_vfp.S
ZTRMMKERNEL = ztrmm_kernel_2x2_vfp.S ZTRMMKERNEL = ztrmm_kernel_2x2_vfp.S
SGEMMKERNEL = sgemm_kernel_4x2_vfp.S SGEMMKERNEL = sgemm_kernel_4x2_vfp.S
SGEMMINCOPY = sgemm_ncopy_4_vfp.S
SGEMMITCOPY = sgemm_tcopy_4_vfp.S
SGEMMINCOPYOBJ = sgemm_incopy.o
SGEMMITCOPYOBJ = sgemm_itcopy.o
SGEMMONCOPY = sgemm_ncopy_2_vfp.S
SGEMMOTCOPY = ../generic/gemm_tcopy_2.c
SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o
DGEMMKERNEL = dgemm_kernel_4x2_vfp.S DGEMMKERNEL = dgemm_kernel_4x2_vfp.S
DGEMMINCOPY = dgemm_ncopy_4_vfp.S
DGEMMITCOPY = dgemm_tcopy_4_vfp.S
DGEMMINCOPYOBJ = dgemm_incopy.o
DGEMMITCOPYOBJ = dgemm_itcopy.o
DGEMMONCOPY = dgemm_ncopy_2_vfp.S
DGEMMOTCOPY = ../generic/gemm_tcopy_2.c
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o
CGEMMKERNEL = cgemm_kernel_2x2_vfp.S CGEMMKERNEL = cgemm_kernel_2x2_vfp.S
CGEMMONCOPY = cgemm_ncopy_2_vfp.S
CGEMMOTCOPY = cgemm_tcopy_2_vfp.S
CGEMMONCOPYOBJ = cgemm_oncopy.o
CGEMMOTCOPYOBJ = cgemm_otcopy.o
ZGEMMKERNEL = zgemm_kernel_2x2_vfp.S ZGEMMKERNEL = zgemm_kernel_2x2_vfp.S
ZGEMMONCOPY = zgemm_ncopy_2_vfp.S
ZGEMMOTCOPY = zgemm_tcopy_2_vfp.S
ZGEMMONCOPYOBJ = zgemm_oncopy.o
ZGEMMOTCOPYOBJ = zgemm_otcopy.o
STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
DTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
DTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
endif

125
kernel/arm/KERNEL.ARMV7
View File

@ -1,86 +1,29 @@
include $(KERNELDIR)/KERNEL.ARMV6
################################################################################# STRMMKERNEL = ../generic/trmmkernel_4x4.c
SAMAXKERNEL = iamax_vfp.S DTRMMKERNEL = ../generic/trmmkernel_4x4.c
DAMAXKERNEL = iamax_vfp.S
CAMAXKERNEL = iamax_vfp.S
ZAMAXKERNEL = iamax_vfp.S
SAMINKERNEL = iamax_vfp.S SGEMMKERNEL = ../generic/gemmkernel_4x4.c
DAMINKERNEL = iamax_vfp.S SGEMMONCOPY = sgemm_ncopy_4_vfp.S
CAMINKERNEL = iamax_vfp.S SGEMMOTCOPY = sgemm_tcopy_4_vfp.S
ZAMINKERNEL = iamax_vfp.S SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o
SMAXKERNEL = iamax_vfp.S DGEMMKERNEL = ../generic/gemmkernel_4x4.c
DMAXKERNEL = iamax_vfp.S DGEMMONCOPY = dgemm_ncopy_4_vfp.S
DGEMMOTCOPY = dgemm_tcopy_4_vfp.S
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o
SMINKERNEL = iamax_vfp.S ifeq ($(ARM_ABI),hard)
DMINKERNEL = iamax_vfp.S
ISAMAXKERNEL = iamax_vfp.S
IDAMAXKERNEL = iamax_vfp.S
ICAMAXKERNEL = iamax_vfp.S
IZAMAXKERNEL = iamax_vfp.S
ISAMINKERNEL = iamax_vfp.S
IDAMINKERNEL = iamax_vfp.S
ICAMINKERNEL = iamax_vfp.S
IZAMINKERNEL = iamax_vfp.S
ISMAXKERNEL = iamax_vfp.S
IDMAXKERNEL = iamax_vfp.S
ISMINKERNEL = iamax_vfp.S
IDMINKERNEL = iamax_vfp.S
SSWAPKERNEL = swap_vfp.S
DSWAPKERNEL = swap_vfp.S
CSWAPKERNEL = swap_vfp.S
ZSWAPKERNEL = swap_vfp.S
SASUMKERNEL = asum_vfp.S
DASUMKERNEL = asum_vfp.S
CASUMKERNEL = asum_vfp.S
ZASUMKERNEL = asum_vfp.S
SAXPYKERNEL = axpy_vfp.S
DAXPYKERNEL = axpy_vfp.S
CAXPYKERNEL = axpy_vfp.S
ZAXPYKERNEL = axpy_vfp.S
SCOPYKERNEL = copy.c
DCOPYKERNEL = copy.c
CCOPYKERNEL = zcopy.c
ZCOPYKERNEL = zcopy.c
SDOTKERNEL = sdot_vfp.S
DDOTKERNEL = ddot_vfp.S
CDOTKERNEL = cdot_vfp.S
ZDOTKERNEL = zdot_vfp.S
SNRM2KERNEL = nrm2_vfpv3.S SNRM2KERNEL = nrm2_vfpv3.S
DNRM2KERNEL = nrm2_vfpv3.S DNRM2KERNEL = nrm2_vfpv3.S
CNRM2KERNEL = nrm2_vfpv3.S CNRM2KERNEL = nrm2_vfpv3.S
ZNRM2KERNEL = nrm2_vfpv3.S ZNRM2KERNEL = nrm2_vfpv3.S
SROTKERNEL = rot_vfp.S
DROTKERNEL = rot_vfp.S
CROTKERNEL = rot_vfp.S
ZROTKERNEL = rot_vfp.S
SSCALKERNEL = scal.c
DSCALKERNEL = scal.c
CSCALKERNEL = zscal.c
ZSCALKERNEL = zscal.c
SGEMVNKERNEL = gemv_n_vfpv3.S SGEMVNKERNEL = gemv_n_vfpv3.S
DGEMVNKERNEL = gemv_n_vfpv3.S DGEMVNKERNEL = gemv_n_vfpv3.S
CGEMVNKERNEL = cgemv_n_vfp.S
ZGEMVNKERNEL = zgemv_n_vfp.S
SGEMVTKERNEL = gemv_t_vfp.S
DGEMVTKERNEL = gemv_t_vfp.S
CGEMVTKERNEL = cgemv_t_vfp.S
ZGEMVTKERNEL = zgemv_t_vfp.S
STRMMKERNEL = strmm_kernel_4x4_vfpv3.S STRMMKERNEL = strmm_kernel_4x4_vfpv3.S
DTRMMKERNEL = dtrmm_kernel_4x4_vfpv3.S DTRMMKERNEL = dtrmm_kernel_4x4_vfpv3.S
@ -88,47 +31,9 @@ CTRMMKERNEL = ctrmm_kernel_2x2_vfpv3.S
ZTRMMKERNEL = ztrmm_kernel_2x2_vfpv3.S ZTRMMKERNEL = ztrmm_kernel_2x2_vfpv3.S
SGEMMKERNEL = sgemm_kernel_4x4_vfpv3.S SGEMMKERNEL = sgemm_kernel_4x4_vfpv3.S
SGEMMONCOPY = sgemm_ncopy_4_vfp.S
SGEMMOTCOPY = sgemm_tcopy_4_vfp.S
SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o
DGEMMKERNEL = dgemm_kernel_4x4_vfpv3.S DGEMMKERNEL = dgemm_kernel_4x4_vfpv3.S
DGEMMONCOPY = dgemm_ncopy_4_vfp.S
DGEMMOTCOPY = dgemm_tcopy_4_vfp.S
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o
CGEMMKERNEL = cgemm_kernel_2x2_vfpv3.S CGEMMKERNEL = cgemm_kernel_2x2_vfpv3.S
CGEMMONCOPY = cgemm_ncopy_2_vfp.S
CGEMMOTCOPY = cgemm_tcopy_2_vfp.S
CGEMMONCOPYOBJ = cgemm_oncopy.o
CGEMMOTCOPYOBJ = cgemm_otcopy.o
ZGEMMKERNEL = zgemm_kernel_2x2_vfpv3.S ZGEMMKERNEL = zgemm_kernel_2x2_vfpv3.S
ZGEMMONCOPY = zgemm_ncopy_2_vfp.S
ZGEMMOTCOPY = zgemm_tcopy_2_vfp.S
ZGEMMONCOPYOBJ = zgemm_oncopy.o
ZGEMMOTCOPYOBJ = zgemm_otcopy.o
STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
DTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
DTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
endif